How to improve computing in matlab?

5 views (last 30 days)
C Zeng
C Zeng on 12 Jun 2012
Hi, it is a general question, in my matlab code I used a lot of loops, and inner loops as well. The result is it cannot get a solution quickly.
I was told that VC is good at loop, but matlab is not, matlab is good at matrix computation. And if I used a lot of matrix calculation, choose matlab.
Is there some suggestions? Thank you!

Accepted Answer

Sean de Wolski
Sean de Wolski on 12 Jun 2012
MATLAB is fine with loops. It's been fine with loops for a very long time but yet still carries that mantra that "loops are bad". They're not. There are some steps you can take to ensure that loops are running at their finest such as:
  • using functions
  • avoiding eval and evil eval wrappers (like str2num (for all you CODY players gaming the system!))
  • preallocating arrays
  • using integers as loop variables when they are used as an index
  • avoid repeated computations
  • avoid changing the class of a variable if possible
  • etc.
(I'm sure Jan will be able to extend this list 10fold).
Moral of the story is: Don't avoid loops like the plague, avoid the bad things above like the plague.
Jan on 13 Jun 2012
@Sean: The list includes the most important strategies already. Some further ideas:
  • Process numerical data columnwise, because accessing the neighboring elements of the memory allows a faster caching. Examples:
X = rand(1e4);
tic; Y = sum(sum(X, 2), 1); toc % 0.19 sec, main work along the rows
tic; Y = sum(sum(X, 1), 2); toc % 0.17 sec, main work along the columns
  • Reduce the number of arithmetic operations when dealing with arrays and scalars:X = rand(1e4); a = 5;Slow: Y = X + pi + 3 + a; This is (((X + pi) + 3) + a) => Three matrix operationsFast: Y = X + (pi + 3 + a); => One matrix operation
  • Use PARFOR to employ more cores.
  • FIND(STRCMP) is much faster than STRMATCH.
  • Do not spend hours for optimizing code until it is so ugly, that you cannot debug it anymore, if it save some seconds of runtime only.
  • Do not let the output of PROFILE confuse you: PROFILE disables the JIT-accelerations, such that Matlab looses its power to process loops efficiently. TIC/TOC is better to measure speed of loops.

Sign in to comment.

More Answers (1)

C Zeng
C Zeng on 17 Jun 2012
thanks to all of your help!


Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!